HASWELL-E 8Core/16T How much faster will it be, because of DDR4 rather than DDR3 ?

[SOFTengCOMPelec]

Title "HASWELL-E 8Core/16T How much faster will it be, because of DDR4 rather than DDR3 ?", says it all.

Any discussion about Haswell-E and/or DDR4 is fine, in this thread.

But the poll is for DDR4's speed improvement, over what it would have been if DDR3, SPECIFICALLY/EXCLUSIVELY for the HASWELL-E 8 Core/16T Enthusiasts Intel cpu
Definition for poll: Speed up based on running a mixture of Benchmarks (all inc games) and real life software usage (compared with if Haswell-E 8C had used DDR3)

I.e. How much faster is a "Haswell-E 8C DDR4", over a (non existent, as will never support DDR3) "Haswell-E 8C DDR3"


E.g. if it was a 10% speed up, and the identical (Haswell-E 8C) took exactly 100.00 Sec, with DDR3
Then DDR4 would take about 91.00 Sec

------------------------------------------------------

DDR4 seems to offer point to point topology.

Advantages
Possibly faster
Lower voltage, so may reduce power consumption
More future safe
Point to point, probably improves things
I think only one slot needs to be filled, no longer need to do it in pairs, but filling all slots, is probably best

Disadvantages
Predicted to be expensive compared to DDR3, in the early stages of DDR4 availability
On a Socket2011++ board, only 4 sockets, rather than 8
Half number of sockets, so need twice capacity Rams, in order to achieve the same (total motherboard) capacity, which tends to be disproportionately more expensive. (e.g. 8 slot x 4 Gb ($20) = $160 for 32 Gb tot, but 4 slot x 8 Gb ($60) = $240 for 32 Gb tot)
May be difficult to get hold of, when Haswell-E first becomes available

Note to mods: I felt it best to put this in the cpu section, as it's about Haswell-E's possible speed benefit because it now uses DDR4. But, if you think this is best in memory, because of DDR4, then please move this thread.

Wiki

 

[SOFTengCOMPelec]

When I recently searched round the web, trying to answer questions like the above, I just kept on getting differing answers, and technical terms that don't make a lot of sense to me, as regards, how much benefit it will give to PC's.
Some articles seem to say no speed improvement, for early DDR4, some say up to about 20%, and other source(s) say up to 50%.
They do seem to somewhat agree, that only certain software applications will really get a speed benefit, and those applications are not common.
Technical terms, such as "point to point", although I partly understand what it means, I can't see either way, how it would alter the performance of a Haswell-E 8 Core/16 Thread, based PC.

 

[CHADBOGA]

With first round of DDR4, I doubt we will see much of a gain over DDR3.

 

[SPBHM]

probably almost nothing I think,

Haswell-e uses quad channel memory anyway, have a huge amount of l3 cache, and DDR3 can go as high as 2133, while DDR4 will start at around the same clock...


DDR4 should be good for IGPs, when they can sell higher clocked (like 3GHz+) modules.

 

[ShintaiDK]

DDR4 is mainly good for IGPs where the bandwidth bottleneck is.

 

[SOFTengCOMPelec]

With first round of DDR4, I doubt we will see much of a gain over DDR3.

probably almost nothing I think,

It is getting very difficult for enthusiasts to improve the (cpu/processing) speed of their rigs (assuming they already have 3770(K) or high end AMD or similar), short of using liquid Nitrogen.

Even if DDR3 ==>> DDR4 was predicted to give a 10% speed gain, beyond other speedups, it would make the (probably) big initial premium price of DDR4 worth it, for enthusiasts determined to get max speed at any cost.

But, if it gives almost no speed improvement, early on, then that makes it much harder to stomach its (probably) high launch price (DDR4), and gives less incentive to go Haswell-E, in the first place.

DDR4 is mainly good for IGPs where the bandwidth bottleneck is.

Which for Haswell-E (no IGPs), is a complete waste of time (from an IGP point of view).

 

[ShintaiDK]

Which for Haswell-E (no IGPs), is a complete waste of time (from an IGP point of view).

But server wise it also offers higher densities and reduced power consumption.

The E series is just a bastard series that lives on borrowed time.

 

[SOFTengCOMPelec]

But server wise it also offers higher densities and reduced power consumption.

Which for Haswell-E (no IGPs), is a complete waste of time (from an IGP point of view).

Sorry, I was not clear enough. I did not mean DDR4 is a waste of time, I meant that DDR4's ability to improve IGP with extra bandwidth, was a waste of time since Haswell-E has no IGP.

The E series is just a bastard series that lives on borrowed time.

Well, for some people, having a real 8 core processor, with 16 threads, good cache, and the latest Haswell IPC improvements, would be their dream computer, especially if they regularly run software, which would benefit from it.

But if Skylake has up to 6 cores, with 12 threads (speculation, as we don't know, when 6 cores are coming to mainstream), some/many people would wait for the (likely to be) considerably cheaper Skylake, and ignore Haswell-E.

 

[ShintaiDK]

But if skylake has up to 6 cores, with 12 threads (speculation, as we don't know, when 6 cores are coming to mainstream), and continues using (probably) cheaper DDR3, some/many people would wait for the (likely to be) considerably cheaper Skylake, and ignore Haswell-E.

Skylake is DDR4. Haswell Xeons get DDR4 to help increase volume and reduce price before hitting mainstream.

 

[SOFTengCOMPelec]

Skylake is DDR4. Haswell Xeons get DDR4 to help increase volume and reduce price before hitting mainstream.

Thanks, I've corrected my earlier post (deleted DDR3 memory, from it).

I bet server owners will be delighted to pay the huge initial DDR4 premium prices, to pay the development cost, so that mainstream can get cheaper DDR4.

Maybe it should be called "Robin Hood memory".


--------------------------------------------

I've just remembered, when I was reading up on DDR4, apparently, it will have built in CRC32 (or something) error checking, which is not as comprehensive like ECC, but does perform some checks, on the address transfers, or something. So, maybe it will be more computationally reliable, than DDR3.

 

[Vesku]

When I recently searched round the web, trying to answer questions like the above, I just kept on getting differing answers, and technical terms that don't make a lot of sense to me, as regards, how much benefit it will give to PC's.
Some articles seem to say no speed improvement, for early DDR4, some say up to about 20%, and other source(s) say up to 50%.
They do seem to somewhat agree, that only certain software applications will really get a speed benefit, and those applications are not common.
Technical terms, such as "point to point", although I partly understand what it means, I can't see either way, how it would alter the performance of a Haswell-E 8 Core/16 Thread, based PC.

I came across similar projections, seems that initial DDR4 will be released near the top of current DDR3 speeds. Expect DDR4 to introduce at 2133 - 2400.

 

[ShintaiDK]

Thanks, I've corrected my earlier post (deleted DDR3 memory, from it).

I bet server owners will be delighted to pay the huge initial DDR4 premium prices, to pay the development cost, so that mainstream can get cheaper DDR4.

Maybe it should be called "Robin Hood memory".


--------------------------------------------

I've just remembered, when I was reading up on DDR4, apparently, it will have built in CRC32 (or something) error checking, which is not as comprehensive like ECC, but does perform some checks, on the address transfers, or something. So, maybe it will be more computationally reliable, than DDR3.

Who says its a huge premium? And did you calculate the TCO savings?

 

[Sable]

Using the massive wealth of knowledge open to me I voted 50% because wild speculation is fun.

 

[SOFTengCOMPelec]

Who says its a huge premium? And did you calculate the TCO savings?

Lots of places say it's going to be more expensive, but the dates of the sources varies, and it is all SPECULATION at this point in time.

If it follows the launch prices of DDR, DDR2 and DDR3 (which I've already forgotten), it surely will be more expensive than current DDR3 prices.

"TCO savings" ... But that is for servers, and people buying Haswell-E's, probably don't care about TCO.
But at a guess, the 20% .. 40% power saving (DDR3 ==>> DDR4), probably gives a TCO (Total Cost of Ownership) advantage (I mean a small saving, NOT a reason to make it worth paying a fortune to use DDR4 over DDR3), so that is a good point.

Source1

There's no rush, really, as Garber notes that the price for DDR4 is outrageous.

 

[SOFTengCOMPelec]

I came across similar projections, seems that initial DDR4 will be released near the top of current DDR3 speeds. Expect DDR4 to introduce at 2133 - 2400.

I was hoping that its new "point to point" mechanism and other improvements, would lead to higher software speeds, even if DDR3=2133, and DDR4=2133.

Because, at the moment, higher DDR3 speeds, make almost no difference to the real life software speeds, at least for non-server, chips and uses (except AMD APU, which can benefit from 1866, rather than 1600, if your APU is fast enough).

EDIT: I was hoping to get 2 speed benefits, from DDR4.
(1)...Instead of going from 1600 to 2400 giving very little, or almost no performance difference, I was hoping that it would give a reasonably significant software execution speed improvement (at least for some software).

(2)...I hoped that it would allow more parallel accesses (which is what I hoped "point to point" meant), also giving a software execution speed up, of a significant amount (at least for some software).

 

[ShintaiDK]

Lots of places say it's going to be more expensive, but the dates of the sources varies, and it is all SPECULATION at this point in time.

If it follows the launch prices of DDR, DDR2 and DDR3 (which I've already forgotten), it surely will be more expensive than current DDR3 prices.

"TCO savings" ... But that is for servers, and people buying Haswell-E's, probably don't care about TCO.
But at a guess, the 20% .. 40% power saving (DDR3 ==>> DDR4), probably gives a TCO (Total Cost of Ownership) advantage (I mean a small saving, NOT a reason to make it worth paying a fortune to use DDR4 over DDR3), so that is a good point.

Source1

There is over a year till you know what DDR4 will actually cost when you will be able to use it in a system. And you said it would be a huge premium, not just a premium.

 

[SOFTengCOMPelec]

There is over a year till you know what DDR4 will actually cost when you will be able to use it in a system. And you said it would be a huge premium, not just a premium.

In one or more of my earlier posts, I used the wording, on the lines of

stomach its (probably) high launch price (DDR4)

I forgot the '(probably)' bit on that last post, sorry.

Yes, it is completely wild speculation, on what the DDR4 memory prices are going to be.

 

[Sequences123]

I was hoping that its new "point to point" mechanism and other improvements, would lead to higher software speeds, even if DDR3=2133, and DDR4=2133.

I doubt that's the case. Your plain old software would benefit much more from being either rewritten or recompiled than from a slight memory speed or better topology. Very few applications out there (either free or commercially) not built for high performance will benefit from this.

I'd venture to guess that a lot of people are already using pseudo point-to-point topology in their builds - dual channel setups where only 1 DIMM per channel is used.

 

[SOFTengCOMPelec]

I doubt that's the case. Your plain old software would benefit much more from being either rewritten or recompiled than from a slight memory speed or better topology. Very few applications out there (either free or commercially) not built for high performance will benefit from this.

I'd venture to guess that a lot of people are already using pseudo point-to-point topology in their builds - dual channel setups where only 1 DIMM per channel is used.

As this thread has progressed, and I have repeatedly read through it, I've painfully had a similar feeling.

Someone else on these forums (possibly Idontcare) used a very powerful analogy, shamelessly repeated here :-

So, what has happened is that some time ago 'all the low lying fruit' has been picked, as regards DDR#-PC memory. Later (but still many, many years ago), the 'medium height fruit was picked off', which tasted somewhat sour and bitter.

So, when we go to DDR4 (no, I DON'T have links, this is my impression from reading this thread), using 'fire engine triple height ladders', powerful drain cleaner, to un-bitter the fruit, an electron microscope to find the highest level fruit, and 6 months to process the fruit, we may see a 0.00000000000000001% speed improvement, if the wind is blowing in the right direction, there is a triple 'ZZZ' in the month, and the moon is SQUARE shaped with pink diamond shaped dots, tonight.

Additionally, from your post, we need to recompile the software.

E.g. GCC -O3 -EnableRichEnthusiastDDR4Speedup -DisableConstantRequestsForLinks -LinuxIsBest -NoMoreJokes

TL;DR;
I was hoping that DDR4 would represent a major break through in memory performance, resulting in nice, 10% speed improvements (to some benchmarks), over equivalent DDR3.

 

[Charles Kozierok]

Ignoring the GPU issue... has there ever been a review posted, anywhere, showing a significant difference in performance between two machines based on slight differences in RAM speed or technology?

I can't recall one, even going back to when Anand started this site. Even when we were arguing over EDO versus the then-barely-new SDRAM, there wasn't a great deal of difference. And that was when CPUs and caches were far more primitive.

RAM speed could be the single most overrated spec in computing.

 

[SOFTengCOMPelec]

RAM speed could be the single most overrated spec in computing.

Yes, very true.
Processor RAM spec (not size, but speed specs), usually makes hardly any difference, except going single channel (when dual is available) DOES seem to slow things down.

With caches, there can be about a 5% .. 10% (maybe up to 20%) speedup, with everything else equal, between slow or less or missing Cache, depending on what the processor is doing.
Video cards DO seem significantly improved, by faster, better forms of memory. Both when you play them, and benchmarks.

I remember in the 'old days', when you could buy an Intel processor (Celeron) for about $50, or so, and hugely overclock it to comparable performance to the top Pentium (circa) $500 processor of the time.

The main spec difference, was that the memory bus speed was considerably less (forgotten figures, but it was something like 66.666 MHz, instead of 133.333 MHz), but in almost all software, the bus/memory speed made almost no difference (or not much) at all, when the final ratio calculated clock speeds were IDENTICAL.

Some people will hate me for saying this, but if Haswell-E effectively goes from a (3770/4770 etc) 4 core to 8 core, apart from some software which utilizes more than 4 cores, it would give even enthusiasts, very little real extra speed, apart from what is today, somewhat rare (in my experience), 8 core utilizing software.
Even software which does go up to 8 cores, it does not necessarily give double the speed of 4 core, as there are limits to how efficiently software can use that many cores.

EDIT: On the other hand, if you are into gaming, and the upcoming new gaming consoles (with 8 cores) mean that much PC gaming software, really needs 8 real cores to get the best out of it, then maybe 8 core processors (such as Haswell-E), would be best for some upcoming games, over the next few years.

But this is of course, wild speculation, as I think some sources think that the extra cores may be used for other functionality (like doing stuff in the background, while you are gaming on the games console), rather than using all 8 cores for a particular game.

 

[podspi]

I vote no difference. Modern processors do a good job at managing the memory bottleneck, and my understanding is that early DDR4 memory is going to be very similar in throughput to high-end DDR3.

But you have to start somewhere.

 

[SOFTengCOMPelec]

I vote no difference. Modern processors do a good job at managing the memory bottleneck, and my understanding is that early DDR4 memory is going to be very similar in throughput to high-end DDR3.

But you have to start somewhere.

That's true (you have to start somewhere).

Yes, a strong pattern/trend of other people's posts seems to be emerging here.

TL;DR;
Modern day caches are so phenomenally optimized, and DDR3 also are well down the high optimization path, that further improvements, are going to show very little practical speed improvement (to benchmarks, and actual real life software execution speeds).
Especially on the early DDR4's.

 

[Ajay]

There is over a year till you know what DDR4 will actually cost when you will be able to use it in a system. And you said it would be a huge premium, not just a premium.

I thing he's roughly correct when it comes to first availability (later this year?) then it will slowly drop till 2015. When DDR4 goes mainstream (Skylake), the price should drop pretty fast (unless DRAM costs are going up in general).

Oh, and I imagine that we might see some higher speed grades in time for Skylake, but that's just a guess (Corsair, et al will want to cash in on the new Desktop CPU).

 

[JoeRambo]

I voted no significant increase, but that is strictly for "workstation" use, as quad channel DDR3 already offers MASSIVE bandwidth and 6 or 8 cores are not enough to chew it in real world tasks that are run on those workstations.

For servers DDR4 is gonna be great, for example 2 socket servers:

1) Currently server DDR3 is stuck on 1600, and if you go beyond 16 slots speed drops fast. While IVB based server processors will increase that to 1866, same 16 slot max speed limitation will remain. So if DDR4 will start @ 2133 and allow filling 24 slots at that speed it can offer substantial advantage in density and speed.
2) Once IVB-EP is out, 12 core / 24 T CPU can get bottlenecked by memory, you don't have to run anything special, enterprise app running on JVM will do, as garbage collectors scale almost perfectly with threads and memory bandwidth.